Early prediction of clinical deterioration in older adults with suspected infection admitted from the emergency department to general wards: a multicenter machine learning study

Older adults with suspected infection admitted from the emergency department (ED) to general wards remain at considerable risk of early clinical deterioration. Conventional early warning scores may perform suboptimally in this population. We aimed to develop and externally validate machine learning models to predict early clinical deterioration within 48 h among older adults with suspected infection. This multicenter retrospective study used electronic medical records from two tertiary medical centers for model development and internal validation between 2016 and 2023, with external validation using data from a regional hospital. Eligible patients were aged 65 years or older with clinically suspected infection. All records from the same patient were assigned to a single partition. Predictors included demographics, comorbidities, laboratory results, and vital signs recorded at emergency department triage and before ward transfer; missing data were handled using MissForest imputation with missingness indicators. Logistic regression (LR), random forest (RF), XGBoost, LightGBM, and CatBoost models were developed and compared with conventional early warning scores (NEWS, MEWS, mSOFA). The primary discrimination metric was the area under the precision-recall curve (AUPRC), with the area under the receiver operating characteristic curve (AUROC), calibration, decision-curve analysis, and Shapley Additive Explanations as secondary analyses. The development cohort comprised 96,871 ED visits, and the external validation cohort comprised 35,361 ED visits. CatBoost achieved the highest AUROC among all evaluated models (internal AUROC, 0.851; external AUROC, 0.829) and significantly outperformed the conventional early warning scores (external AUROC ≤ 0.70; all DeLong p < 0.001). For precision–recall performance (external event-rate baseline, 0.017), CatBoost achieved the highest external AUPRC (0.134), approximately eight times the baseline and more than three times the best conventional score, with the other machine-learning models ranging from 0.093 to 0.119. Calibration curves demonstrated good agreement between predicted and observed risks for the gradient-boosting models, whereas logistic regression showed greater deviation from the ideal diagonal. Gradient boosting models, particularly CatBoost, achieved good discrimination for early clinical deterioration in older adults with suspected infection admitted from the ED to general wards, and outperformed conventional early warning scores. These models may support pre-ward-transfer risk assessment at the time of the ward-admission decision; prospective studies are warranted to confirm real-world effectiveness and guide clinical implementation.